The 1.5% lifetime prevalence estimate of probable NAP based on preliminary clinical review of CID open-ended responses is in the middle of the range of prevalence estimates in past community epidemiological surveys (Bland et al 1988
; Canino et al 1987
; Hwu et al 1989
; Keith et al 1991
; Kendler et al 1996
; Lee et al 1990
; Wells et al 1989
; Wittchen et al 1992
). This was achieved, though, using a much smaller set of screening questions than in previous surveys as well as with a much smaller proportion of respondents who endorsed the questions than in previous surveys (6 compared to 14 screening questions and 9.3% compared to 28.4% of respondents who endorsed screening questions in the NCS-R compared to the NCS). Only a tiny proportion of open-ended responses were classified as the respondent misunderstood the question, compared to a majority in previous surveys. This shows that the NCS-R screening question revision improved on the screens used in previous surveys.
The 0.3% lifetime prevalence estimate of clinician-diagnosed NAP in the NCS-R is considerably lower than the 1.5% estimate based on preliminary clinical review of the CIDI responses. This difference is less than in previous community epidemiological surveys, though, providing another indication that the NCS-R CIDI NAP symptom screening questions improved on those used in previous surveys. The larger discrepancy in previous surveys is due to the vast majority of self-reported psychotic symptoms in response to fully structured symptom questions representing either experiences that can easily be misinterpreted by researchers as delusions (e.g., reports of being followed) or misunderstandings of survey questions about hallucinations (e.g., reports of having excellent vision). Systematic clinical evaluations of respondents with positive responses to such questions in earlier surveys have shown only a small minority to be truly psychotic (Hanssen et al 2003
; Johns et al 2004
; Kendler et al 1996
). It is consequently important to be cautious in interpreting the comparatively high prevalence estimates of psychotic spectrum experiences found using fully-structured assessments in recent community epidemiological surveys (Johns et al 2004
; Maric et al 2003
; van Os et al 2001
NCS-R findings regarding the correlates of NAP are also consistent with previous surveys and incidence studies. These include the findings that NAP has a median age-of-onset in the late teens or early twenties that is somewhat earlier among men than women (Jablensky et al 1992
), that NAP is significantly related to disadvantaged social status (Jablensky 2000
), that NAP is highly comorbid with a wide range of other mental disorders (Kendler et al., 1996
) and substance use disorders (Murray et al 2002
), that NAP is associated with substantial impairment in all areas of life (Jablensky et al 1980
), and that the vast majority of people with NAP are in contact with the treatment system (Sartorius et al 1986
). We also found a somewhat higher relative-odds of NAP among men than women (1.6), which is consistent with a modest, but statistically significant, elevation among men compared to women found in meta-analysis of incidence studies (McGrath et al 2004
There are also two noteworthy NCS-R findings that either diverge from the results of previous surveys or that raise concerns about the current results. The first is that the median age-of-onset of NAP among men (13) and the low end of the IQR of that estimate (10), as assessed by the CIDI, are both lower than in previous studies. These discrepancies could reflect nothing more than the imprecision of sub-sample estimates based on the small number of NAP cases detected in the NCS-R. The second finding that warrants comment along the same lines is that over one-fourth of respondents diagnosed with lifetime NAP in the SCID met lifetime hierarchy-free criteria for bipolar disorder in the CIDI. Although this pattern could lead to questions about the accuracy of the SCID diagnoses, more detailed case-by-case review of complete case records confirmed that NAP was the more appropriate diagnosis than bipolar disorder in all these cases.
Despite the general consistency with previous research, the NCS-R results have to be interpreted in the context of three limitations. First, NAP was assessed comprehensively only among respondents with clinical reappraisal interviews. MI was used for the remainder. Concern about this limitation is reduced by the AUC of the imputation equation being high, which means that statistical power is close to what it would have been if NAP were assessed with the SCID for all respondents. The MI method adjusts for the imprecision introduced by imputation, with prevalence estimated without bias.
Second, a much larger sample than the 2322 Part II NCS-R respondents screened for NAP is required to study such a rare disorder powerfully. Implications of this limitation are seen in –, where large ORs are often not significant. Caution is needed in interpreting results because of this limitation.
Third, people with NAP might have been under-represented in the NCS-R due to any of three factors: (i) The sampling frame excluded population segments (non-English speakers, institutions, homeless) that might have a high NAP prevalence. (ii) NCS-R non-respondents might have a higher prevalence of NAP than respondents. (iii) Some respondents who screened negative for NAP in the CIDI might actually have NAP.
Each of these three is plausible. Studies of people in non-household populations (e.g., homeless, prison, nursing home) document higher prevalence of NAP than in the household population (Fischer and Breakey 1991
; Keith et al 1991
). This does not introduce much bias into total-population estimates, though, due to the small proportion of the population not in the English-speaking household population. Kendler et al. (1996)
estimated that this exclusion leads to no more than a 0.1% downward bias in the estimated lifetime total population prevalence of NAP.
Under-representation of household residents with NAP is potentially much more important. Survey response rates have declined steadily for the past several decades (de Leeuw and de Heer 2002
) and are now often as low as 50–60% (Groves and Couper 1998
). With such low response rates, NAP prevalence as low as 2% among non-respondents would lead to 50% under-estimation of prevalence. We were unable to investigate this in the NCS-R non-response survey (Kessler et al 2004
), as in previous non-response studies (Badawi et al 1999
; Eaton et al 1992
; Kessler et al 1995
) due to the rarity of NAP. However, a Swedish study found that people with a history of treated schizophrenia based on registry data were significantly less likely than others to participate in a mental health survey (Allgulander 1989
NAP screening measures having NPV less than 1.0 could have an even more dramatic effect on estimated prevalence. At least some confirmed cases of NAP are subsequently misclassified even in clinical follow-up studies (Tsuang et al 1981
). Even a very small decrement in NPV, such as from 1.00 to .99, would more than quadruple the estimated lifetime prevalence of NAP in the NCS-R. This possibility can be evaluated with SCID clinical reappraisal interviews in a probability sub-sample of screened negatives. However, the number of screened negatives would have to be very large to be helpful due to the low prevalence of NAP. For example, in order to estimate prevalence in a plausible range (i.e., 0.0–0.3%) with even minimal precision (i.e., a standard error equal to half the range), SCID clinical reappraisal interviews would have to be administered to at least 1000 screened negatives. This was not possible for financial reasons in the NCS-R.
The problem of non-response bias is even more difficult to address. A methodological study was carried out in the NCS-R where limited information was obtained from a probability sub-sample of non-respondents. However, only a minority of non-respondents agreed to participate in that survey even with a substantial financial incentive. Non-respondents with NAP could well have been less likely to participate than other non-respondents in light of common paranoid symptoms in NAP.
Based on these concerns, we should consider whether our NAP prevalence estimate is consistent with other sources of data. A number of NAP total-population incidence studies have been carried out using clinical methods to assess all incident cases in a population over some time interval (Jablensky et al 1992
; Sartorius et al 1986
). These results can be used to generate alternative estimates of NAP lifetime risk. Jablensky (2000)
reviewed these studies. Annual NAP incidence estimates were 0.016–0.042%. In equilibrium, these estimates can be multiplied by number of birth cohorts at risk to estimate lifetime risk. Assuming conservatively that the main age range of risk is between the ages of 15 and 55 (recognizing that risk is not constant in that age range), estimated lifetime risk would be 0.64–1.68%. This is considerably higher than the adjusted 0.1–0.8% survey estimate. Estimated lifetime risk should be higher than estimated lifetime (to date) prevalence, but this should only play a small part in the discrepancy due to the early age-of-onset distribution of NAP.A much more plausible explanation for the lower survey estimates is downwardly biased because of the problems reviewed above.
Two of these three problems could be resolved with larger surveys and larger clinical re-interview sub-samples and expanded sampling frames to include homeless and institutionalized people. However, these design changes would not resolve the most serious problem – the presumed high non-response rate among people with NAP. One way to finesse this problem would be to change the focus to a more tractable research task, as in the Study on Low Prevalence (i.e., psychotic) Disorders (SLPD) carried out in conjunction with the Australian National Survey of Mental Health and Wellbeing. The SLPD carried out a screen for psychosis among all patients in treatment for mental health problems in four areas of Australia in a specified month and then carried out an in-depth assessment of NAP with screened positives (Jablensky 2000
). An exercise of this sort is tractable and useful in estimating demand for treatment of psychosis even if it does not estimate unmet need for such treatment.
Realistic prospects for resolving the non-response problem to allow NAP prevalence to be estimated with reasonable precision and accuracy are less clear. Survey sampling specialists have developed several methods to estimate prevalence and correlates of rare behaviors (Kalton and Anderson 1986
), but none of them is likely to be effective in dealing with the NAP non-response problem. The one possible exception is the multiplicity sampling method (Sudman et al 1988
), in which informants report on rare behaviors of well-defined networks (e.g., their first-degree relatives), where probability of hearing about each detected case is calculated based on a reconstruction of the size of the case’s network, and an underestimation adjustment weight is used to estimate prevalence based on network size. Family studies show that informants provide information on between 27% and 60% of independently confirmed psychotic cases (Andreasen et al 1986
; Davies et al 1997
; Roy et al 1996
), making multiplicity sampling potentially feasible. However, probability of informant reports is significantly related to characteristics of the informant and the case as well as density of psychosis in the network (Roy et al 1996
). This means that it would be complicated to reconstruct probability of detection for each confirmed case from informant reports. Despite these difficulties, though, multiplicity sampling should be carefully considered if future general population studies attempt to estimate NAP prevalence.